CN115329821A - Ship noise identification method based on pairing coding network and comparison learning - Google Patents

Abstract

The invention provides a ship noise identification method based on a pairing coding network and comparison learning, and belongs to the field of ship radiation noise identification. Aiming at the problem of disguise of target radiation noise on a spectrogram, a convolutional neural network and a long-time memory network are set up and mapped twice by utilizing an unsupervised learning thought based on comparison on the premise of ensuring the real-time effect of a model. In addition, the invention uses the time sequence matrix and the cross entropy function to force the model to discard the original acoustic useless information and noise of the sample, and learn the high-level characteristic difference between different targets, thereby improving the accuracy in identifying the targets. According to the method, through comparison of the learning idea and reconstruction of the loss function, the simple structure of the model is ensured, the effect of high real-time identification accuracy is achieved, and the problem that the ship target pretends to the spectrogram of the ship target and interferes with the existing identification system is effectively solved.

Description

Ship noise identification method based on pairing coding network and comparison learning

Technical Field

The invention relates to the field of underwater acoustic signal processing, in particular to a ship noise identification method based on pairing coding network and comparison learning.

Background

With the rapid development of deep learning in the fields of computer vision, pattern recognition, acoustic signal processing and the like, other research ideas are provided for target recognition of passive sonar signals, and the deep learning becomes an important research topic. However, the accuracy of the model is inevitably reduced after the number of classification targets is increased. On the other hand, the deep learning model needs a large amount of sample training, the signal-to-noise ratio of signal data in water is low, high-quality samples are deficient, and data in different water areas are difficult to migrate and use. The method solves the very challenging problem of insufficient data in the field, arouses the research interest of a plurality of researchers, and further provides small sample learning and zero sample learning.

The underwater target identification refers to a process of generating effective audio signals by processing signals based on signals collected by a sonar and further identifying the signals. In the military field, research and development of underwater target individual identification technology have extremely important significance. However, due to the complexity and variability of underwater environments (such as acoustic medium constraints and heterogeneity), and the underwater signal acquisition and transmission processing costs, the underwater available data is often not accurate enough, so how to improve the accuracy and rapidity of detection and identification of underwater targets, and reduce the calculation and communication costs becomes a problem to be solved urgently.

The knowledge obtained by learning in the source task is used for helping the learning of the new task target in the transfer learning, and a good training effect can be achieved when the training data of the new task is less. On the basis of transfer learning, aiming at the problem of insufficient data, the small sample learning is suitable for the learning scene with deficient samples caused by objective reasons. Zero sample learning is in a test set, some classes are not in a training set, and a model is trained by using samples of the training set, so that the model can be applied to the test set to correctly identify labels which do not exist in the training set. The method aims to solve the problems of the appearance of a new target and the repeated appearance of the same target in a short period of time, and the small sample learning and the zero sample learning are both unavailable, so that the method becomes a research hotspot at present. The technology for realizing the underwater acoustic individual recognition architecture design mainly comprises three steps. The first step is that a neural network is trained through a large number of obtained samples, and the difficulty of intelligent identification under the conditions of changing hydrological environment and few samples for training is broken through; secondly, extracting time-frequency information to mine multi-dimensional features with rich individual differences; and thirdly, learning individual characteristics by using a deep learning network to achieve the purpose of classification and identification. However, in the field of underwater target identification, the underwater data acquisition requires huge economic investment, and most of the existing related data sets have high confidentiality and large regional characteristics. In this case, it is difficult to achieve good results by means of the conventional machine learning method driven by large data. Therefore, small sample learning is required to build unsupervised, end-to-end classification models from these rare image data and to improve accuracy in model identification.

Disclosure of Invention

Aiming at the problems that in the prior art, the ship radiation noise identification accuracy is low and the noise reduction means needs to be adjusted and optimized manually, the invention provides a ship noise identification method based on a pairing coding network and contrast learning.

In order to achieve the purpose, the invention adopts the technical scheme that:

a ship noise identification method based on pairing coding network and contrast learning comprises the following steps:

step 1, dividing a long-time ship noise signal received by a hydrophone into signal samples with fixed time length, mapping the signal samples to a feature space through a convolutional neural network, and obtaining a first feature tensor of each signal sample;

step 2, constructing a cyclic neural network formed by a long-time and short-time memory layer, and sequentially inputting the first feature tensor obtained in the step 1 into the cyclic neural network according to the time sequence of the signal sample to obtain a corresponding second feature tensor; the second feature tensor of the previous moment is stored in the long-time and short-time memory layer and is fused in the second feature tensor of the later moment;

step 3, constructing a loss function in a cross entropy form, and calculating a loss value according to the first feature tensor and the second feature tensor;

step 4, constructing a first optimizer and a second optimizer, wherein the first optimizer optimizes the convolutional neural network, the second optimizer optimizes the combination of the convolutional neural network and the cyclic neural network, and each optimizer updates the network parameters of the corresponding neural network by using the loss value calculated in the step 3 through a gradient back propagation method to obtain the trained convolutional neural network and the trained cyclic neural network;

and 5, inputting the ship noise signal to be identified into the trained convolutional neural network to obtain a corresponding first characteristic tensor, flattening the first characteristic tensor into a vector, and then obtaining a classification result through the full connection layer and the softmax classifier to complete the identification of the ship noise signal.

Further, the specific manner of step 1 is as follows:

step 101, dividing a long-time ship noise signal received by a hydrophone into 3.2768 seconds, namely 16384 point number signal samples, adjusting the dimension of the signal samples from 1 × 16384 to 128 × 128, and reading the signal samples into a convolutional neural network according to 64 one-batch number;

step 102, mapping the signal samples to a dimensional space of 256 feature points through a 8*8 convolution kernel with a step length of 4, and obtaining a feature matrix with a size of 256 × 15 through a nonlinear activation function ReLU and a maximum pooling layer with a size of 2;

and 103, passing the feature matrix obtained in the step 102 through 4*4, a convolution kernel with the step size of 2, the ReLU and the maximum pooling layer to obtain 256 × 3 feature tensors.

Further, the specific manner of step 3 is as follows:

(1) Constructing a diagonal tensor of a corresponding dimension according to the first characteristic tensor and the second characteristic tensor;

(2) Reconstructing the diagonal tensor into a matrix shifted to the right by one bit, and copying according to the number of samples to obtain a time sequence tag array;

(3) Performing matrix multiplication on the first feature tensor and the second feature tensor, and keeping the number of samples unchanged to obtain uniform sample features;

(4) Batch cross entropy log loss was calculated for the unified sample feature and time series tag array.

The invention has the beneficial effects that:

1. the invention provides a novel ship noise identification method by combining the idea of comparative learning and a pairing coding network. By constructing an improved feature extraction module, the purpose of learning high-dimensional information behind the radiation noise of the ship is achieved.

2. According to the invention, a long-time and short-time memory layer (GRU) is constructed, the structure can extract the information of the data, and the sample is subjected to nonlinear mapping again, so that the problem of time sequence correlation of the acoustic data is effectively solved.

3. The method constructs a loss function in a cross entropy form, uniformly optimizes the identification part and the noise reduction part, forces the model to discard original acoustic useless information and noise of a sample through two times of nonlinear mapping, and learns high-level characteristic differences of different targets. Therefore, the capability of separating noise of the network is improved while the network identification capability is ensured.

In a word, the method avoids the limitation of the traditional spectrum analysis through an end-to-end method, and enables the passive ship identification task according to the radiation noise to achieve higher accuracy and stronger robustness through a plurality of optimization technologies.

Drawings

FIG. 1 is a general framework diagram of an embodiment of the invention;

FIG. 2 is a schematic structural diagram of a feature extraction module and a timing information fusion module according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a calculation of a loss function based on comparative learning according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.

A ship noise identification method based on pairing coding network and comparison learning comprises the following steps:

s1, feature extraction: carrying out first mapping on the signal through a feature extraction module;

s2, fusing time sequence information: fusing the output of the step S1 with the time information before the fusion through a time sequence information fusion module;

s3, calculating a loss function: calculating loss according to the characteristics obtained after the processing of the steps S1 and S2;

s4, model training and optimizing: after the gradient is returned through the SGD, parameters of the feature extraction module and the time sequence information fusion module are updated to obtain a trained feature extraction module;

s5, model application: and classifying the ship noise by using the trained feature extraction module and combining a softmax classifier, and completing the identification.

The characteristic extraction module is a CNN (convolutional neural network), and the timing information fusion module is an RNN (recurrent neural network) formed by a GRU (long-time memory layer).

Wherein, step S4 includes the following steps:

s41: constructing a first SGD optimizer, and returning the loss calculated in the step S3 to the time sequence information fusion module through a gradient back propagation algorithm;

s42: and constructing a second SGD optimizer, returning the loss obtained by calculation in the step S3 to the feature extraction module through a gradient back propagation algorithm, and mixing the loss with the loss returned by the first SGD optimizer, so as to optimize the combination of the feature extraction module and the time sequence information fusion module.

The following is a more specific example:

referring to fig. 1, a ship noise identification method based on pairing coding network and contrast learning includes the following steps:

step S1, feature extraction:

in the embodiment, in order to learn high-dimensional information (such as tail rotor and engine generating different noises) behind ship radiation noise, a feature extraction module as shown in fig. 2 is built.

The specific way of feature extraction is as follows:

s11: the method comprises the steps of dividing a long-time-period ship noise signal with sampling frequency of 5000Hz received by a hydrophone into 3.2768 seconds (16384 points) signal samples, and then mapping the signal samples to a feature space through a Convolutional Neural Network (CNN).

Adjust the sample dimension from 1 × 16384 to 128 × 128 to facilitate the two-dimensional convolution process and read the data set into the model as many as 64 one lot;

s12: mapping 128 × 128 samples into a dimension space of 256 feature points through a convolution kernel of 8*8 with a step size of 4, and then passing through a nonlinear activation function ReLU and a maximum pooling layer with the size of 2;

wherein, the two-dimensional convolution formula is as follows:

in the formula, w _m，n Is a convolution kernel, x _i+m，j+n Is the current value of the sample characteristic, a _i，j Is a mapped value, w _b Is a preset bias value of the convolution layer;

the ReLU formula is as follows:

f(x)＝max(0，x)

in the formula, x is the current sample characteristic, and f (x) is a new sample characteristic obtained by comparing x with 0 and then taking the maximum value.

The maximum pooling layer is used to maximize the values in the matrix of 2*2, as follows:

in the formula, N _i Is the abscissa, C _j Is the ordinate, stride [0 ]]X h + m and stride [1]X w + n denotes each block region after the original sample is divided into different regions. By taking the maximum value in the transverse direction and the maximum value in the longitudinal direction of each block region, the methodThe maximum value of the whole area is used as the representation of the current area and the input of the next layer.

S13: and (3) passing the feature matrix with the size of 256 × 15 obtained in the step S12 through 4*4, a convolution kernel with the step size of 2, the ReLU and the maximum pooling layer to obtain the feature tensor of 256 × 3.

S2, fusing time sequence information:

considering the time sequence correlation of the acoustic data, the invention constructs a long-time and short-time memory layer (GRU) shown in FIG. 2 to extract the preamble information of the data, and non-linearly maps the sample again; the specific mode is as follows:

s21: adjusting the dimensionality of the feature tensor obtained in the step S13 to 9 × 256, and initializing one feature tensor for each time sequence;

s22: and inputting the feature tensor obtained in the step S13 into the GRU layer of 256 × 256, combining the previous information extracted at the previous moment, outputting the feature tensor of 9 × 256, and storing the information at the current moment in the GRU layer for use in the next calculation. The correlation formula is as follows:

r _t ＝σ(W _ir x _t +b _ir +W _hr h _(t-1) +b _hr )

z _t ＝σ(W _iz x _t +b _iz +W _hz h _(t-1) +b _hz )

n _t ＝tanh(W _in x _t +b _in +r _t *(W _hn h _(t-1) +b _hn ))

h _t ＝(1-z _t )*n _t +z _t *h _(t-1)

in the formula, W _ir 、W _hr 、W _iz 、W _hz 、W _in 、W _hn 、b _ir 、b _hr 、b _iz 、b _hz 、b _in 、b _hn Is the weight, x, of the hidden layer of the GRU _t Is the current sample, r _t Is the weight of the last moment information, z _t Is the total amount of data that can be transferred at the last time, n _t Is the calculated output of the current time information, h _(t-1) Is the output of the layer above the whole GRU layer，h _t Is the actual output of the entire GRU layer.

Step S3, calculating a loss function:

considering how to directly perform target identification from ship radiation noise, a proper mapping mode needs to be found, and an original sample is mapped to a hidden space, so that the distance between positive and negative samples is short, the network identification capability is ensured, and the noise separation capability of the network is improved. As shown in FIG. 3, the method constructs a loss function in a cross entropy form, optimizes the identification part and the noise reduction part uniformly, forces the model to discard original acoustic useless information and noise of the sample through two times of nonlinear mapping, and learns high-level feature differences of different targets. In fig. 3, x1 and x2 represent a sample at the moment and a sample at the next moment of the input radiation noise signal, y1 and y2 represent sample representations of x1 and x2 obtained through a first non-linear mapping (i.e., feature extraction), and z1 and z2 represent sample representations of x1 and x2 obtained through a second non-linear mapping (i.e., time sequence information fusion).

The specific mode is as follows:

s31: initializing a diagonal matrix (9*9) with the same time sequence dimension as the characteristic tensor output by the S22, converting the diagonal matrix to obtain a time sequence matrix which pays attention to the current time and the previous time, and copying 64 times (the number of batch samples);

s32: performing matrix dot multiplication in batches on the feature tensor obtained in the step S13 and the feature tensor obtained in the step S22;

in the formula, bmat1 is the feature tensor obtained in S13, bmat2 is the feature tensor obtained in S22, and out is obtained as an output by performing dot multiplication on a matrix corresponding to the dimension.

S33: calculating the intra-batch logarithmic cross entropy loss function for the tensor and timing matrix of S31:

l _n，c ＝-w _n，c [p _c y _n，c ·logσ(x _n，c )+(1-y _n，c )·log(1-σ(x _n，c ))]

in the formula, w _n，c Is a weight matrix, p _c Is a penalty factor for classification errors, y _n，c Is the corresponding label in the timing matrix, x _n，c Is the corresponding value in the tensor, σ is the activation function (sigmoid).

S4, model training optimization:

when the loss function is calculated in S33, the gradients of the two modules are separated by taking the logarithm, and become an additive relationship, so that two optimizers are used to optimize the two modules respectively.

The specific mode is as follows:

s41: and (4) constructing two SGD optimizers, and updating parameters of the time sequence fusion module and the feature extraction module by using the loss obtained in the step (S3) through a gradient back propagation algorithm. The correlation formula is as follows:

x _t+1 ＝x _t -η _t g _t

in the formula, x _t Is the current value of the parameter, x _t+1 Is the parameter value of the next moment, η _t Is the learning rate, which in this example takes the value of 0.0001,g _t Is an updated value calculated according to loss;

expression to x _t Partial differential on the independent variable, E.]Is shown for g _t The desired mean is calculated.

In the formula, X ₀ And X ₁ Respectively the updated values returned by the two optimizers, lambda is the weight, the value is 0.5 in the example, and the updated values are obtained by adding

S5, model verification and application:

after the model training is completed, the feature extraction module is applied to an actual task, and further processing is needed by using a classification module. When the model effect is tested, verifying the model classification effect by adopting a K mean value clustering algorithm; during actual ship radiation noise identification, a classifier is adopted for supervised training to finely adjust model parameters, and a better identification effect is achieved.

The specific mode is as follows:

s51: and in the process of model testing, 4-mean clustering is carried out on the characteristic matrix obtained in the step S1. 4 samples are randomly selected as the mass center, each sample is iteratively divided into the nearest mass center after the Euclidean distance is judged, and the mass center is recalculated until the mass center is not changed any more. The calculation method is as follows:

in the formula, x _i Is the ith sample, x _iu Is the u-th characteristic value of the i-th sample, n is the total number of the characteristic values, and the Euclidean distance dist is obtained _ed (x _i ，x _j ) And then, iteratively calculating the centroid.

Through calculation, the accuracy rate of distinguishing each type of sample is 100%, and the classification recognition effect of the model is verified;

s52: flattening the characteristic matrix obtained in the step (1) into vectors, then outputting classification results through 2304 × 4 full connection layers and softmax. In practical application, a small amount of data of known labels are used for carrying out supervised training on the classifier and the feature extraction module, and parameters are finely adjusted, so that the model is more suitable for practical tasks. The formula of the softmax classifier is as follows:

where exp (.) represents the natural logarithm of the solution, f _c Representing a vector labeled C, which represents the number of types.The softmax function maps the input vector to real numbers between 0-1 and normalizes the guaranteed sum to 1 so that the sum of the probabilities for multiple classes is also exactly 1. Thus, the output of the softmax function is the probability of each class being taken.

In a word, aiming at the problem of disguise of target radiation noise on a spectrogram, the method utilizes an unsupervised learning idea based on comparison to establish a convolutional neural network and a long-time memory network and perform mapping twice on the premise of ensuring the real-time effect of a model. In addition, the time sequence matrix and the cross entropy function are used, so that the model is forced to discard original acoustic useless information and noise of the sample, high-level feature differences among different targets are learned, and the accuracy in target identification is improved. According to the method, through comparison of the learning thought and reconstruction of the loss function, the effect of high real-time identification accuracy is achieved while the simple structure of the model is ensured, and the problem that the ship target disguises the spectrogram of the ship target and interferes the existing identification system is effectively solved.

It should be noted that the above-mentioned are only specific embodiments of the present invention, and any feature disclosed in the present specification may be replaced by other equivalent or alternative features having similar purpose, unless otherwise specified; all of the disclosed features, or all of the method or process steps, may be combined in any combination, except mutually exclusive features and/or steps.

Claims (3)

1. A ship noise identification method based on pairing coding network and comparison learning is characterized by comprising the following steps:

step 1, dividing a long-time ship noise signal received by a hydrophone into signal samples with fixed time length, mapping the signal samples to a feature space through a convolutional neural network, and obtaining a first feature tensor of each signal sample;

step 2, constructing a cyclic neural network formed by a long-time and short-time memory layer, and sequentially inputting the first feature tensor obtained in the step 1 into the cyclic neural network according to the time sequence of the signal sample to obtain a corresponding second feature tensor; the second feature tensor of the previous moment is stored in the long-time and short-time memory layer and is fused with the second feature tensor of the next moment;

step 3, constructing a loss function in a cross entropy form, and calculating a loss value according to the first characteristic tensor and the second characteristic tensor;

step 4, a first optimizer and a second optimizer are constructed, the first optimizer optimizes the convolutional neural network, the second optimizer optimizes the combination of the convolutional neural network and the cyclic neural network, and each optimizer updates network parameters of the corresponding neural network through a gradient back propagation method by using the loss value calculated in the step 3 to obtain the trained convolutional neural network and the trained cyclic neural network;

and 5, inputting the ship noise signal to be identified into the trained convolutional neural network to obtain a corresponding first characteristic tensor, flattening the first characteristic tensor into a vector, and then obtaining a classification result through the full connection layer and the softmax classifier to complete the identification of the ship noise signal.

2. The ship noise identification method based on the pairing coding network and the comparative learning according to claim 1, wherein the specific mode of the step 1 is as follows:

step 101, dividing a long-time ship noise signal received by a hydrophone into 3.2768 seconds, namely signal samples with 16384 point numbers, adjusting the dimension of the signal samples from 1 × 16384 to 128 × 128, and reading the signal samples into a convolutional neural network according to 64 signals in a batch;

step 102, mapping the signal samples to a dimensional space of 256 feature points through a 8*8 convolution kernel with a step length of 4, and obtaining a feature matrix with a size of 256 × 15 through a nonlinear activation function ReLU and a maximum pooling layer with a size of 2;

and 103, passing the feature matrix obtained in the step 102 through 4*4, a convolution kernel with the step size of 2, the ReLU and the maximum pooling layer to obtain 256 × 3 feature tensors.

3. The ship noise identification method based on pairing coding network and comparative learning according to claim 1, wherein the specific mode of step 3 is as follows:

(1) Constructing a diagonal tensor of a corresponding dimension according to the first characteristic tensor and the second characteristic tensor;

(2) Reconstructing the diagonal tensor into a matrix which is shifted to the right by one bit, and copying according to the number of samples to obtain a time sequence tag array;

(3) Performing matrix multiplication on the first feature tensor and the second feature tensor, and keeping the number of samples unchanged to obtain uniform sample features;

(4) Batch cross entropy log loss was calculated for the unified sample feature and time series tag array.

Images